A weighted twin support vector machine as a potential discriminant analysis tool and evaluation of its performance for near-infrared spectroscopic discrimination of the geographical origins of diverse agricultural products

Abstract

A weighted twin support vector machine (wTWSVM) was proposed as a potential discriminant analysis tool and its utility was evaluated for near-infrared (NIR) spectroscopic identification of the geographical origins of 12 different agricultural products including black soybean and garlic. In the wTWSVM, weights were applied on each variable in the sample spectra to highlight detailed NIR spectral features and the optimal weights to minimize the discrimination error were iteratively searched. Then, the weighted spectra were employed to determine the samples’ geographical origins using a TWSVM adopting two non-parallel hyperplanes for the discrimination. For the performance evaluation, SVM, TWSVM, and wTWSVM were separately used for the two-group discriminations and their accuracies were comparatively analyzed. When the SVM and TWSVM accuracies were compared, the improvements by using the TWSVM were significant (95% confidence level) for 10 out of the 12 products. Moreover, the accuracy improvements with the wTWSVM against SVM were significant for all the 12 products. In the case of the TWSVM-wTWSVM accuracy comparison, the improvements by the wTWSVM were also significant for 10 products, thereby demonstrating superior discrimination performance of wTWSVM. Based on the overall results, the wTWSVM could be a potential chemometric tool for discriminant analysis and expandable to other areas such as spectroscopy-based biomedical disease diagnosis and forensic analysis. © 2021 Elsevier B.V.